Anthropic is opening up its AI text detection system to third-party developers through a new API, making it possible to check whether a piece of writing likely came from Claude. The company is using a variant of SynthID Text, the watermarking method Google DeepMind published in Nature last year. The technique works by subtly shifting the randomness source during word selection, leaving a traceable statistical pattern behind. Anthropic claims this doesn’t affect “the content, level of creativity, or readability of Claude’s text.”
The catch? The watermark has real limits. Short texts and fact-heavy passages where Claude has few alternative phrasings to choose from won’t carry a reliable signal. Code is similarly unreliable. If a human rewrites generated text heavily enough, the watermark can vanish entirely. And the system can only flag that Claude was likely involved — it can’t distinguish between full generation and heavy editing, nor can it tell whether text came from a different AI model or a human.
This rollout is being driven by regulation rather than market demand. Anthropic joined roughly 190 signatories of the EU Code of Practice on transparency for AI-generated content, which includes watermarking obligations under the EU AI Act. Since there’s no practical way to restrict the feature by region, the company is launching it worldwide. Models released after August 2, 2025 support watermarking out of the box, with older Claude versions getting the feature in the coming months. For files, Anthropic uses the C2PA open standard to attach metadata without altering the content itself.
What makes this genuinely interesting is how it differs from existing AI detection tools like Pangram. Those services don’t have access to Anthropic’s internal systems. They scan for stylistic fingerprints — repeated phrases, predictable sentence structures, telltale word choices. That approach has been notoriously unreliable, producing false positives and easy to evade with light editing. Watermarking is a fundamentally different beast: it’s a cryptographic-style signal baked into the generation process itself, not a guess based on vibes. If the implementation holds up in practice, it could shift the entire AI detection conversation away from probabilistic guessing toward verifiable provenance. The open question is whether users will accept invisible watermarks on everything Claude produces, and whether other major labs follow suit or wait to see how the market reacts.
